Land cover and its dynamics significantly influence many environmental processes and properties. Land cover change, is for instance, a major component of the global carbon cycle and a significant control on biodiversity. Although the significance of land cover is recognised our knowledge on its precise effects is limited by the poor quality of land cover data, particularly at regional to global scales.
The only feasible method to acquire land cover data at such scales is through satellite remote sensing. Typically land cover data are derived from remotely sensed data through the application of a supervised image classification. This approach has not, however, always provided sufficiently accurate or appropriate land cover data sets. A fundamental problem is that the conventional methods used are 'hard' techniques, in which each image pixel is associated with only one class throughout the classification. Although this may be reasonable with some relatively fine spatial resolution remotely sensed data sets, coarse spatial resolution satellite sensor imagery such as that acquired by the NOAA AVHRR (finest spatial resolution of 1.1km), which are the backbone of the regional-global scales investigations, are usually dominated by pixels of mixed land cover composition. Failure to accommodate for mixed pixels will result in a poor representation of land cover distribution and incorrect estimates of class extent derived from it. Errors may then feed into subsequent analysis. For instance, estimates of the carbon flux to the atmosphere associated with deforestation may be incorrect due to misclassification errors associated with mixed pixels.
There are two main sources of mixed pixels. First, the land cover classes on the ground may not lie in a discrete mosaic on the surface of the Earth but instead intergrade gradually. Indeed the boundary between two classes is often vague but may be important as an indicator of environmental change. Since the intergradation of classes implies their spatial co-existence, conventional 'hard' image classification techniques will be unable to provide an appropriate representation of the gradient between two classes and their inferred boundary. Softer/fuzzy approaches designed to accommodate fuzziness and the presence of mixed pixels may, however, be able to model the class gradients and enable the location of the boundary to be inferred more accurately than a conventional 'hard' classification. Second, even if the classes could be considered as being discrete, the pixels of coarse spatial resolution imagery will often straddle the boundary of two or more classes; clearly the proportion of mixed pixels arising from this effect is a function of the sensor's spatial resolution and the fabric of the landscape. These land cover classes may be more appropriately represented by fraction images which depict the proportional land cover composition of each pixel rather than a 'hard' classification output. Irrespective of their mode of formation, mixed pixels cannot be sensibly included at any point in the conventional classification process and as they generally dominate remotely sensed data sets alternative approaches are required. Here the use of a fuzzy classification achieved with the application of an artificial neural network is discussed. The aim is to use the neural network to derive estimates of the sub-pixel land cover composition. These may then be mapped or used for the derivation of estimates of the spatial extent of land cover classes over a region. The paper reports on an investigation aimed at mapping land cover from NOAA AVHRR data of West Africa. Here land cover classes that may be considered to be discrete (e.g. water) and continuous (e.g. the forest-savanna transition) are encountered. The aim was to derive a method to represent the land cover that could reflect the gradient from dense tropical forest to tropical savanna that may be used to monitor its seasonal dynamics and derive accurate estimates of the extent of land cover classes. Emphasis is placed on three issues. First, a description of the method for unmixing sub-pixel land cover using an artificial neural network will be presented. The method is based on a standard feedforward network architecture; throughout relatively simple networks containing a single and small hidden layer were used to obtain a high generalisation capacity. Unlike other unmixing methods it will be stressed that the neural network based approach has several major advantages, notably the independence from assumptions about the nature of the mixing process and freedom from a requirement for often elusive end-member spectra. Second, the ability to represent the transition zone between the forest and savanna classes and its seasonal variation will be assessed. It will be shown that the seasonal movements of this transition may be identified and characterised. Particular attention will be paid to the behaviour of the neural network output units for pixels from the transition zone between the forest and savanna, where, in terms of class membership, there is maximum fuzziness. Third, the accuracy with which the extent of land cover class may be estimated from the fuzzy classification will be assessed. This will be illustrated with reference to the estimation of tropical forest extent and discussed in relation to the derivation of deforestation estimates required for studies of the carbon cycle and species-areas curves used in biodiversity assessments. The accuracy of these estimates will be expressed relative to forest inventory ground data.
The paper will conclude with a brief discussion on further refinements and possibilities of artificial neural networks for image classification.